The sample of observed defaults significantly understates the average firm's true expected cost of default due to a sample selection bias. To quantify this selection bias, I use a dynamic capital structure model to estimate firm-specific expected default costs. The average firm expects to lose 45% of firm value in default, a cost higher than existing estimates. However, the average cost among defaulted firms in the estimated model is only 25%, a value consistent with existing empirical estimates from observed defaults. This substantial selection bias helps to reconcile the levels of leverage and default costs observed in the data.